Dataset Preparation

This document shows how to convert ocr annotation to the general format (not including LMDB) for model training.

You may also refer to convert_datasets.sh which is a quick solution for converting annotation files of all datasets under a given directory.

To download and convert OCR datasets to the required data format, please refer to these instructions.

• Born-Digital Images
• CASIA-10K
• CCPD
• Chinese text recognition
• COCO-Text
• CTW
• ICDAR2015
• ICDAR2019 ArT
• LSVT
• MLT2017
• MSRA-TD500
• MTWI-2018
• RCTW-17
• ReCTS
• SCUT-CTW1500
• SROIE
• SVT
• SynText150k
• SynthText
• TextOCR
• Total-Text

Text Detection/Spotting Annotation

The format of the converted annotation file should follow:

img_61.jpg\t[{"transcription": "MASA", "points": [[310, 104], [416, 141], [418, 216], [312, 179]]}, {...}]

Taking ICDAR2015 (ic15) dataset as an example, to convert the ic15 dataset to the required format, please run

# convert training anotation
python tools/dataset_converters/convert.py \
        --dataset_name  ic15 \
        --task det \
        --image_dir /path/to/ic15/det/train/ch4_training_images \
        --label_dir /path/to/ic15/det/train/ch4_training_localization_transcription_gt \
        --output_path /path/to/ic15/det/train/det_gt.txt

# convert testing anotation
python tools/dataset_converters/convert.py \
        --dataset_name  ic15 \
        --task det \
        --image_dir /path/to/ic15/det/test/ch4_test_images \
        --label_dir /path/to/ic15/det/test/ch4_test_localization_transcription_gt \
        --output_path /path/to/ic15/det/test/det_gt.txt

Text Recognition Annotation

Common Dataset Format

The annotation format for text recognition dataset follows

word_7.png  fusionopolis
word_8.png  fusionopolis
word_9.png  Reserve
word_10.png CAUTION
word_11.png citi

Note that image name and text label are seperated by \t.

To convert, please run:

# convert training anotation
python tools/dataset_converters/convert.py \
        --dataset_name  ic15 \
        --task rec \
        --label_dir /path/to/ic15/rec/ch4_training_word_images_gt/gt.txt \
        --output_path /path/to/ic15/rec/train/ch4_training_word_images_gt/rec_gt.txt

# convert testing anotation
python tools/dataset_converters/convert.py \
        --dataset_name  ic15 \
        --task rec \
        --label_dir /path/to/ic15/rec/ch4_test_word_images_gt/gt.txt \
        --output_path /path/to/ic15/rec/ch4_test_word_images_gt/rec_gt.txt

LMDB Dataset Format

Some of the dataset can be converted to LMDB format. Currently, this is only supported for the SynthText and SynthAdd datasets.

To convert to LMDB format, please run

python tools/dataset_converters/convert.py \
    --dataset_name synthtext \
    --task rec_lmdb \
    --image_dir /path/to/SynthText \
    --label_dir /path/to/SynthText_gt.mat \
    --output_path ST_full